CoRL Related Work
Robot learning is a confluence field: any given CoRL paper inherits from at least
three literatures moving at different speeds. Reviewers here are typically expert
in one lane and conversant in the rest — so a related-work section that covers
only the authors' home lane reads as blinkered to a third of the pool.
The three lanes, plus two overlays
| Lane |
What reviewers check |
Typical miss |
| Robot-learning lineage (CoRL, RSS, ICRA/IROS learning papers) |
Do you know the last 2–3 years of the task family you claim? |
Citing only your lab's chain of prior work |
| ML methods (NeurIPS/ICLR/ICML) |
Is the algorithmic idea actually new, or a known method re-embodied? |
Renaming an established technique with a robotics word |
| Classical robotics (planning, control, estimation) |
Does a non-learning baseline solve your task? Why learn at all? |
Dismissing model-based methods without citation or comparison |
| Overlay: foundation-model / VLA wave |
Position relative to the current generation of pretrained robot policies |
Comparing against a wave that is two generations stale |
| Overlay: concurrent arXiv |
Preprint culture is aggressive here; overlap appears mid-review |
Silence on a widely known concurrent preprint |
Practical breadth check: your citation list should contain entries from at least
two recent CoRL volumes, at least one ML venue, and at least one classical
robotics source — or an explicit reason why a lane is genuinely empty.
Handling velocity: the field moves during your review
Between the late-May CoRL deadline and September decisions, the arXiv state of
your subfield will change. Norms to apply:
- Work that appeared publicly after your submission cannot be held against
your novelty; if a reviewer raises it, note the dates factually in the rebuttal
and offer a comparison in revision if it is genuinely relevant.
- Work that was on arXiv well before the deadline is fair game even without a
venue stamp — this community treats known preprints as citable context. Sweep
arXiv thoroughly in deadline month; "it wasn't published yet" does not excuse
missing a preprint your reviewers all saw on social media.
- For close concurrent work, add a labeled paragraph ("Concurrent work") that
differentiates on approach and evidence rather than priority claims.
Positioning moves that work at this venue
- Differentiate on evidence regime, not just method. "Prior work
demonstrates X in simulation; we provide the first real-robot evaluation with
a measured transfer gap" is a legitimate and well-received axis here.
- Name the assumption you remove. "Existing approaches assume known object
poses; we operate from raw RGB" — a robotics-meaningful delta beats an
architecture tweak in this pool's eyes.
- Give the task-family history one honest sentence of state-of-the-art,
including the best non-learning result. Erasing the classical solution to make
the learning look necessary is a detectable and damaging move.
- Baselines follow citations: any method you describe as closest related
work should appear in your tables, or the text should say why it cannot
(
corl-experiments covers fairness mechanics).
PMLR citation hygiene — including the year-offset trap
CoRL proceedings publish through PMLR, and two mechanical errors recur:
- The year offset. PMLR volumes sometimes carry a publication year after
the conference year — e.g., the CoRL 2021 best paper (Chen et al.,
in-hand re-orientation) is
PMLR v164, 2022, and SayCan from CoRL 2022
appears with 2023 pagination metadata. Cite the conference edition in prose
("at CoRL 2021") even when the BibTeX year says otherwise, and keep the pair
consistent across your references.
- Venue misattribution. Neighboring PMLR volumes belong to different
conferences; copying a BibTeX entry from a random site routinely mislabels a
CoRL paper as appearing at another venue or vice versa. Resolve every robot
learning citation against the volume index (v205 = CoRL 2022, v229 = 2023,
v270 = 2024, v305 = 2025).
@inproceedings{chen2022inhand,
title = {A System for General In-Hand Object Re-Orientation},
author = {Chen, Tao and Xu, Jie and Agrawal, Pulkit},
booktitle = {Proceedings of the 5th Conference on Robot Learning (CoRL 2021)},
series = {Proceedings of Machine Learning Research},
volume = {164},
pages = {297--307},
year = {2022}, % PMLR publication year != conference year — keep both visible
publisher = {PMLR}
}
Prefer arXiv-to-PMLR upgrade passes before camera-ready: many robot-learning
papers you cited as preprints during writing will have acquired proceedings
entries by October.
Double-anonymous self-citation
- Cite your own prior work in the third person, exactly as you would a
stranger's; the giveaway is not the citation but the possessive framing
("building on our platform [12]").
- If the paper visibly extends a system with a recognizable name, do not
amputate the citation — anonymity through omission breaks the scholarly
record; third-person distance is the sanctioned tool.
- Watch transitive leaks: citing an unreleased internal report, a thesis, or a
dataset only your lab could hold identifies you as surely as a name.
Section-audit checklist
[ ] All three lanes represented (or absence argued), plus VLA-wave currency
[ ] ≥2 recent CoRL volumes cited within the claimed task family
[ ] Best classical/non-learning approach acknowledged, compared or excused
[ ] Concurrent-work paragraph drafted if a known preprint is close
[ ] Every "closest work" citation appears in the experiments, or is excused
[ ] PMLR volume/year pairs verified against volume indexes; no misattribution
[ ] Self-citations third-person; no possessive framing; no transitive leaks
Verify volume anchors and any new proceedings arrangements at
https://proceedings.mlr.press/ and the live corl.org pages; this file's anchors
were checked 2026-07-08.
1---2name: corl-related-work3description: Use when positioning a CoRL paper against the literature — the robot-learning lineage across CoRL/RSS/ICRA, the ML-methods stream from NeurIPS/ICLR/ICML, the fast-moving VLA and foundation-model wave, classical robotics baselines, concurrent arXiv work, and PMLR citation hygiene including the year-offset trap.4---56# CoRL Related Work78Robot learning is a confluence field: any given CoRL paper inherits from at least9three literatures moving at different speeds. Reviewers here are typically expert10in one lane and conversant in the rest — so a related-work section that covers11only the authors' home lane reads as blinkered to a third of the pool.1213## The three lanes, plus two overlays1415| Lane | What reviewers check | Typical miss |16|---|---|---|17| Robot-learning lineage (CoRL, RSS, ICRA/IROS learning papers) | Do you know the last 2–3 years of the *task family* you claim? | Citing only your lab's chain of prior work |18| ML methods (NeurIPS/ICLR/ICML) | Is the algorithmic idea actually new, or a known method re-embodied? | Renaming an established technique with a robotics word |19| Classical robotics (planning, control, estimation) | Does a non-learning baseline solve your task? Why learn at all? | Dismissing model-based methods without citation or comparison |20| Overlay: foundation-model / VLA wave | Position relative to the current generation of pretrained robot policies | Comparing against a wave that is two generations stale |21| Overlay: concurrent arXiv | Preprint culture is aggressive here; overlap appears mid-review | Silence on a widely known concurrent preprint |2223Practical breadth check: your citation list should contain entries from at least24two recent CoRL volumes, at least one ML venue, and at least one classical25robotics source — or an explicit reason why a lane is genuinely empty.2627## Handling velocity: the field moves during your review2829Between the late-May CoRL deadline and September decisions, the arXiv state of30your subfield will change. Norms to apply:3132- Work that appeared publicly *after* your submission cannot be held against33 your novelty; if a reviewer raises it, note the dates factually in the rebuttal34 and offer a comparison in revision if it is genuinely relevant.35- Work that was on arXiv well *before* the deadline is fair game even without a36 venue stamp — this community treats known preprints as citable context. Sweep37 arXiv thoroughly in deadline month; "it wasn't published yet" does not excuse38 missing a preprint your reviewers all saw on social media.39- For close concurrent work, add a labeled paragraph ("Concurrent work") that40 differentiates on approach and evidence rather than priority claims.4142## Positioning moves that work at this venue4344- **Differentiate on evidence regime, not just method.** "Prior work45 demonstrates X in simulation; we provide the first real-robot evaluation with46 a measured transfer gap" is a legitimate and well-received axis here.47- **Name the assumption you remove.** "Existing approaches assume known object48 poses; we operate from raw RGB" — a robotics-meaningful delta beats an49 architecture tweak in this pool's eyes.50- **Give the task-family history one honest sentence of state-of-the-art**,51 including the best non-learning result. Erasing the classical solution to make52 the learning look necessary is a detectable and damaging move.53- **Baselines follow citations**: any method you describe as closest related54 work should appear in your tables, or the text should say why it cannot55 (`corl-experiments` covers fairness mechanics).5657## PMLR citation hygiene — including the year-offset trap5859CoRL proceedings publish through PMLR, and two mechanical errors recur:60611. **The year offset.** PMLR volumes sometimes carry a publication year after62 the conference year — e.g., the CoRL **2021** best paper (Chen et al.,63 in-hand re-orientation) is `PMLR v164, 2022`, and SayCan from CoRL **2022**64 appears with 2023 pagination metadata. Cite the conference edition in prose65 ("at CoRL 2021") even when the BibTeX year says otherwise, and keep the pair66 consistent across your references.672. **Venue misattribution.** Neighboring PMLR volumes belong to different68 conferences; copying a BibTeX entry from a random site routinely mislabels a69 CoRL paper as appearing at another venue or vice versa. Resolve every robot70 learning citation against the volume index (v205 = CoRL 2022, v229 = 2023,71 v270 = 2024, v305 = 2025).7273```bibtex74@inproceedings{chen2022inhand,75 title = {A System for General In-Hand Object Re-Orientation},76 author = {Chen, Tao and Xu, Jie and Agrawal, Pulkit},77 booktitle = {Proceedings of the 5th Conference on Robot Learning (CoRL 2021)},78 series = {Proceedings of Machine Learning Research},79 volume = {164},80 pages = {297--307},81 year = {2022}, % PMLR publication year != conference year — keep both visible82 publisher = {PMLR}83}84```8586Prefer arXiv-to-PMLR upgrade passes before camera-ready: many robot-learning87papers you cited as preprints during writing will have acquired proceedings88entries by October.8990## Double-anonymous self-citation9192- Cite your own prior work in the third person, exactly as you would a93 stranger's; the giveaway is not the citation but the possessive framing94 ("building on our platform [12]").95- If the paper visibly extends a system with a recognizable name, do not96 amputate the citation — anonymity through omission breaks the scholarly97 record; third-person distance is the sanctioned tool.98- Watch transitive leaks: citing an unreleased internal report, a thesis, or a99 dataset only your lab could hold identifies you as surely as a name.100101## Section-audit checklist102103```text104[ ] All three lanes represented (or absence argued), plus VLA-wave currency105[ ] ≥2 recent CoRL volumes cited within the claimed task family106[ ] Best classical/non-learning approach acknowledged, compared or excused107[ ] Concurrent-work paragraph drafted if a known preprint is close108[ ] Every "closest work" citation appears in the experiments, or is excused109[ ] PMLR volume/year pairs verified against volume indexes; no misattribution110[ ] Self-citations third-person; no possessive framing; no transitive leaks111```112113Verify volume anchors and any new proceedings arrangements at114https://proceedings.mlr.press/ and the live corl.org pages; this file's anchors115were checked 2026-07-08.
brycewang-stanford